On 23.01.2008 [19:29:15 +0200], Pekka J Enberg wrote:

Hi,

On Wed, 23 Jan 2008, Mel Gorman wrote:

quoted

Applied in combination with the N_NORMAL_MEMORY revert and it fails to
boot. Console is as follows;

Thanks for testing!

On Wed, 23 Jan 2008, Mel Gorman wrote:

quoted

[c0000000005c3b40] c0000000000dadb4 .cache_grow+0x7c/0x338
[c0000000005c3c00] c0000000000db518 .fallback_alloc+0x1c0/0x224
[c0000000005c3cb0] c0000000000db920 .kmem_cache_alloc+0xe0/0x14c
[c0000000005c3d50] c0000000000dcbd0 .kmem_cache_create+0x230/0x4cc
[c0000000005c3e30] c0000000004c049c .kmem_cache_init+0x1ec/0x51c
[c0000000005c3ee0] c00000000049f8d8 .start_kernel+0x304/0x3fc
[c0000000005c3f90] c000000000008594 .start_here_common+0x54/0xc0

0xc0000000000dadb4 is in cache_grow (mm/slab.c:2782).
2777            local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2778    
2779            /* Take the l3 list lock to change the colour_next on this node */
2780            check_irq_off();
2781            l3 = cachep->nodelists[nodeid];
2782            spin_lock(&l3->list_lock);
2783    
2784            /* Get colour for the slab, and cal the next value. */
2785            offset = l3->colour_next;
2786            l3->colour_next++;

Ok, so it's too early to fallback_alloc() because in kmem_cache_init() we 
do:

        for (i = 0; i < NUM_INIT_LISTS; i++) {
                kmem_list3_init(&initkmem_list3[i]);
                if (i < MAX_NUMNODES)
                        cache_cache.nodelists[i] = NULL;
        }

Fine. But, why are we hitting fallback_alloc() in the first place? It's 
definitely not because of missing ->nodelists as we do:

        cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];

before attempting to set up kmalloc caches. Now, if I understood 
correctly, we're booting off a memoryless node so kmem_getpages() will 
return NULL thus forcing us to fallback_alloc() which is unavailable at 
this point.

As far as I can tell, there are two ways to fix this:

  (1) don't boot off a memoryless node (why are we doing this in the first 
      place?)

On at least one of the machines in question, wasn't it the case that
node 0 had all the memory and node 1 had all the CPUs? In that case, you
would have to boot off a memoryless node? And as long as that is a
physically valid configuration, the kernel should handle it.

  (2) initialize cache_cache.nodelists with initmem_list3 equivalents
      for *each node hat has normal memory*

I am still wondering why this worked before, though.

I bet we didn't notice this breaking because SLUB became the default and
SLAB isn't on in the test.kernel.org testing, for instance. Perhaps we
should add a second set of runs for some of the boxes there to run with
CONFIG_SLAB on?

I'm curious if we know, for sure, of a kernel with CONFIG_SLAB=y that
has booted all of the boxes reporting issues? That is, did they all work
with 2.6.23?

Thanks,
Nish